A Domain Term Mining Method Based on Knowledge Graph Recall

Abstract

A domain term mining method based on knowledge graph recall, comprising the following steps: S1: Extract domain concept words according to the current document; S2: Extract triples according to the context of the domain concept words in the current document and generate a first graph; S3: Point the domain concept words to a prior knowledge graph, and generate a second graph according to the path relationship of the concept words in the knowledge graph; S4: Select nodes with high similarity of graph embedding vectors in the second graph for recall, and add them to the first graph, and generate a third graph by finding known relationships according to the newly added nodes; S5: Construct a prompt text according to the third graph to confirm the definition of the domain concept words; The present invention uses the domain knowledge graph and the general knowledge graph constructed by known concept words in the domain as prior knowledge, and confirms the relationship between concept words by the way of recall in the prior knowledge graph, avoiding misunderstandings caused by the ambiguity of concept term, and improving the accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and more particularly to a method for mining domain terms based on knowledge graph recall. Background Art

[0002] Mining domain terms helps to identify key domain concepts, monitor the development trends of the domain, provide information for decision-makers, and improve communication.

[0003] Currently, there are three mainstream methods for term mining in the industry:

[0004] Rule-based: The rule-based method is a method for identifying terms by establishing a series of predefined rules. These rules need to borrow the knowledge of domain experts to construct rule templates and can be continuously optimized and improved manually. This method is simple and easy to implement, but its portability is poor. Specific rules are obtained by consuming a large amount of manpower, and the effectiveness of the rules is easily interfered.

[0005] Machine learning-based: The machine learning method uses machine learning algorithms to identify terms. Machine learning algorithms can learn the laws related to domain terms through training and automatically identify terms according to these laws. However, this method requires a large amount of labeled data and computing resources, and it takes a long time to train the model.

[0006] Combined method: The combined method is a method that combines the rule-based method and the machine learning method. This method can use the rule-based method to improve the accuracy and reliability of the model, and at the same time, it can also use the machine learning method to expand the data set and process a large number of terms. The advantage of this method is that it can comprehensively utilize the advantages of the two methods and can process more terms. However, this method requires more time and computing resources to train the model and requires more complex data preprocessing.

[0007] The main problems in current domain term mining mainly focus on:

[0008] Strong ambiguity and great disambiguation difficulty: Some terms may have multiple meanings or interpretations, making it difficult to determine their intended uses in a specific environment during the annotation stage and directly affecting the performance of the algorithm on new data.

[0009] Insufficient timeliness and susceptibility to noise: For newly emerging terms in the domain, due to their small appearance in the training samples and the lack of prior knowledge of domain experts, both the recall rate and the precision rate are insufficient, and they are more likely to be affected by irrelevant or misleading data.

[0010] Strong dependence and poor generalization ability: The training samples of domain terms and rule writers are generally restricted by the complexity and diversity of the domain itself. For domains with poor data accumulation, it is difficult for current mainstream methods to have good scalability.

[0011] Therefore, how to improve the accuracy of domain concept terms is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0012] In view of this, the present invention provides a domain term mining method based on knowledge graph recall. A domain knowledge graph is constructed from known concept terms in the domain, and the relationships between concept terms are confirmed by means of recall in the domain knowledge graph, avoiding misunderstandings caused by the ambiguity of concept term terms and improving accuracy.

[0013] In order to achieve the above object, the present invention adopts the following technical solutions:

[0014] A domain term mining method based on knowledge graph recall, comprising the following steps:

[0015] S1: Extract domain concept terms and context from the current document; in order to process multimodal data such as pictures, voices, and videos, ocr technology and voice-to-text technology can be used to first convert the text in pictures, voices in voices and videos into text data.

[0016] S2: Extract triples according to the context of the domain concept terms in the current document and generate a first graph;

[0017] S3: Point the domain concept terms to a prior knowledge graph and generate a second graph according to the path relationship of the concept terms in the knowledge graph;

[0018] S4: Select nodes with high similarity of graph embedding vectors in the second graph for recall and add them to the first graph, and generate a third graph according to the known relationships of the added nodes found in the second graph;

[0019] S5: Construct a prompt text according to the third graph to confirm the definition of the domain concept terms.

[0020] Preferably, the S1 is specifically: training a concept term recognition model for extraction, and the concept term recognition model includes a named entity recognition model and a large language model;

[0021] In the offline training process, the large language model is used for vocabulary annotation, generating marked samples, and recognizing the annotated samples; counting the confidence of each annotated vocabulary and filtering the annotated vocabulary through a threshold; for parameter fine-tuning according to internal domain documents to develop prompt words;

[0022] Use the large language model to predict external general documents and extract high-frequency co-occurring words; match according to the high-frequency co-occurring words to obtain external labeled samples;

[0023] The named entity recognition model is used to train according to the external labeled samples.

[0024] Preferably, the steps of the concept word recognition model for online model inference include:

[0025] S11: Perform a preliminary recall on the current document through the named entity recognition model, and mine document keywords to form a keyword table;

[0026] S12: Adjust the keyword table through the large language model to obtain the domain concept words.

[0027] Preferably, the specific content of S2 includes: The large language model takes the extracted concept words as entities, confirms the entity relationships according to the context of the concept words in the current document, extracts triples, and constructs a first graph.

[0028] Preferably, the specific content of S3 includes: The large language model chains the entities in the first graph to an existing concept graph or a general knowledge graph according to the entity relationships in the first graph;

[0029] Extract the linked entities and entity relationships in the existing concept graph or general knowledge graph to generate a second graph.

[0030] Preferably, the training steps of the large language model further include:

[0031] Construct an entity disambiguation training set, which includes sample sentences and sample entities in the sample sentences; construct a domain knowledge graph and a general knowledge graph, and generate knowledge graph labels through entity alignment; the knowledge graph labels are provided with target semantic entities and non-target semantic entities corresponding to the sample entities;

[0032] Use the entity disambiguation training set to train the large language model; the trained large language model is used to obtain the target semantic entity corresponding to the sample entity among multiple candidate semantic entities in the knowledge graph.

[0033] Preferably, S3 further includes: judging the graph similarity of the generated second graph, and when the similarity meets the preset threshold, execute S4.

[0034] Preferably, the calculation method of the graph similarity is:

[0035] sim = k1 * node_sim(g1, g2) + k2 * edge_sim(g1, g2) + k3 * group_sim(g1, g2)

[0036] Among them, sim represents the similarity eigenvalue between the first knowledge graph g1 and the second knowledge graph g2. node_sim is to take out the graph embedding vectors of all the nodes in g1 and g2 and calculate the average value of the cosine similarities respectively; edge_sim is to take out the graph embedding vectors of all the edges in g1 and g2 and calculate the average value of the cosine similarities respectively; group_sim is to take 3-step distances of all the nodes in g1 in the knowledge graph, and the similarity obtained by dividing the number of intersection points of all the nodes and all the nodes in g2 by the number of nodes in g2. k1, k2, and k3 are weight constants.

[0037] Preferably, S4 specifically includes:

[0038] Calculate the distance similarity between the domain concept word and all entities in the second knowledge graph;

[0039] Set the distance similarity threshold;

[0040] When the distance similarity threshold is satisfied and the corresponding domain concept word does not hit the first knowledge graph, select the corresponding entity of the domain concept word in the second knowledge graph to recall the first knowledge graph and add it to the first knowledge graph, and generate a third knowledge graph according to the newly added nodes to find the known relationships.

[0041] Preferably, S5 is specifically:

[0042] Construct a paraphrased text according to the third knowledge graph to represent the relationships between the nodes, and combine the meta-information of the current document, the known definition information of the domain concept word, and the context text of the domain concept word to form a prompt text;

[0043] When the domain concept word is a new word, confirm the relationship between the new word and the domain words in the library according to the prompt text, and summarize to generate a definition abstract of the new word.

[0044] It can be seen from the above technical solutions that compared with the prior art, the beneficial effects of the domain term mining method based on knowledge graph recall provided by the present invention are as follows:

[0045] 1. Construct a domain knowledge graph from the known concept words in the domain, and confirm the relationships between the concept words by means of recall in the domain knowledge graph, avoiding misunderstandings caused by the ambiguity of concept word terms and improving the accuracy.

[0046] 2. Existing domain term mining devices often require a large amount of manual adjustment and human intervention, which is a long and time-consuming process. By automatically annotating based on large language models to form prompt words to guide the recognition of domain concept words, and using a named entity recognition model to achieve the mining of new words, machine learning and natural language processing technologies are utilized, enabling these tasks to be completed faster while providing more accurate results.

[0047] 3. The present invention can process various types of data, including pictures, videos, etc. Knowledge framework pictures often contain a lot of domain knowledge structure information. The present invention describes the picture information and then extracts the domain terms and relationship sets from it, making the information richer and more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0049] Figure 1 Schematic diagram of a domain term mining method based on knowledge graph recall provided by the present invention.

[0050] Figure 2 Schematic diagram of the offline training process of the concept word recognition model in the embodiments of the present invention.

[0051] Figure 3 Schematic diagram of the online inference process of the concept word recognition model in the embodiments of the present invention.

[0052] Figure 4 Online inference process for new domain term relationship recognition and summary in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0054] As Figure 1 , the embodiments of the present invention disclose a domain term mining method based on knowledge graph recall, including the following steps:

[0055] S1: Extract domain concept words and their contexts from the current document; to process multi-modal data such as images, voices, and videos, OCR technology and voice-to-text technology can be used to first convert the text in images, voices in voices, and sounds in videos into text data.

[0056] S2: Extract triples based on the context of the domain concept words in the current document and generate the first knowledge graph.

[0057] S3: Link the domain concept words to the prior knowledge graph and generate the second knowledge graph according to the path relationship of the concept words in the knowledge graph.

[0058] S4: Select nodes with high similarity of graph embedding vectors in the second knowledge graph for recall and add them to the first knowledge graph. Generate the third knowledge graph according to the known relationships of the newly added nodes found in the second knowledge graph.

[0059] S5: Construct a prompt text based on the third knowledge graph to confirm the definition of the domain concept words.

[0060] In this embodiment, the present invention can confirm the relationships of entities in the current document in the prior knowledge graph through linking, and through the method of entity recall, realize the interpretation of entities, that is, concept words, in the current document by prior knowledge, avoiding misunderstandings caused by concept word ambiguity.

[0061] In addition, if there are new words in the current document and corresponding entities cannot be found in the prior knowledge through linking, but the relationship between the prior knowledge and the new words can be indirectly confirmed using other entities in the current document, thus realizing the definition of the new words.

[0062] In one embodiment, S1 specifically includes: learning the concept word features in the domain through training a concept word recognition model and finally using it to recognize the document to extract the concept words in the document.

[0063] To further implement the above technical solution, the present invention provides a concept word recognition model, which includes a named entity recognition model and a large language model.

[0064] Such as Figure 2 , during the offline training process, the large language model is used for vocabulary annotation, recognizing the annotation samples, counting the confidence of each annotated vocabulary, and filtering through a threshold.

[0065] When performing the vocabulary annotation task, the input information received by the large language model includes task instructions and descriptions, confirming the professional field and marking symbols, such as "in the computer field, use # to mark the start position of the term and | to mark the end position of the term". Finally, perform the marking task according to the instruction requirements, as follows:

[0066] Input sample sentence: "For example, using large language models to solve the common chicken and rabbit cage problem in mathematics. In the early stage, CoT provided a small number of Q&A examples, which is called few-shot prompting technology."

[0067] Output the marked sample: "For example, using #large language model| to solve the common chicken and rabbit cage problem in mathematics. In the early stage, #CoT| provided a small number of Q&A examples, which is called #few-shot prompting| technology."

[0068] When identifying the marked sample, the large language model discriminates each labeled word one by one to determine whether it is a domain concept word and outputs the probability that it is a domain concept word. The examples are as follows:

[0069] Input: Is "hallucination" in the sentence "Wikipedia + large model beats hallucination! Stanford WikiChat outperforms GPT-4 with an accuracy rate of up to 97.3%" a computer term?

[0070] Output: 1.0

[0071] Input: Is "hallucination" in the sentence "I slept too little, resulting in a short-term hallucination." a computer term?

[0072] Output: 0.0

[0073] Input: Is "n" in the sentence "n = n + 1, where n is a variable name" a computer term?

[0074] Output: 0.1

[0075] In the recognition task, the labeled words can be used to form a prompt template to guide the model recognition, enabling the large language model to recognize more types of domain words and improving the recognition accuracy.

[0076] After completing the recognition task, count each labeled word, calculate the average score of the words, and set the threshold as the average score for filtering.

[0077] Furthermore, based on the large language model, the present invention introduces a named entity recognition model to capture newly emerging terms in the field, so as to enrich the prior knowledge of training samples. Specifically, a large number of Internet documents are collected through the Octopus tool, and the inference of concept word recognition is carried out using the above-mentioned large language model service based on the prompt template. Since the Internet samples used are external general documents, new concept word terms and new semantics of known concept words will appear, and thus a large number of concept word results with slightly poor accuracy will be generated. The remotely supervised method is used to screen for newly emerging terms in these results and generate external annotation samples. The samples obtained in this way are used to fine-tune a small model (traditional language model or large language model with a small number of parameters), that is, the named entity recognition model. This process needs to be continuously iterated to maintain the effect, so that the effect of the small model is close to or even exceeds that of the large model, and finally replaces the large model to reduce the online inference cost. The remotely supervised method mainly obtains a set of concept words with high-frequency co-occurrence relationships from specific domain documents, and directly generates external annotation samples if they match in the sample sentences.

[0078] In this embodiment, as Figure 3 , when the trained concept word recognition model performs online inference, the steps include:

[0079] S11: Mining keywords from the current document through the named entity recognition model to form a keyword table;

[0080] S12: Supplementing and adjusting the keyword table through the large language model to obtain the domain concept words.

[0081] Specifically, the named entity recognition model can learn the patterns of specific types of entity words through training. When encountering new words, it can infer the boundaries and types of entity words based on their context information and the co-occurrence relationship between words. The large language model has strong semantic understanding ability and instruction following ability. First, prompt engineering is used to guide the large language model to label the domain words in the text to supplement the word table obtained by the named entity recognition model. Then, the large language model is used as a discriminator to further screen the labeled vocabulary, understand its semantics based on the context of the current text where the keyword is located, and discriminates one by one whether it is a domain concept word. Based on the above supplementation and adjustment of the large language model, the recall rate and accuracy of domain word recognition are improved.

[0082] To further implement the above technical solution, the prior knowledge graph in S2 includes: a known domain knowledge graph and a general knowledge graph. Among them, the domain knowledge graph is a knowledge graph constructed based on known domain terms in the target domain and the relationships between them, and the general knowledge graph is a general knowledge graph with a wider coverage such as Wikidata. The domain knowledge graph can be regarded as a subset of the general knowledge graph. Some open-source domain knowledge graphs can correspond to the general knowledge graph, but the general knowledge graph often cannot cover all the knowledge in a specific domain.

[0083] Furthermore, the domain knowledge graph and the general knowledge graph can be entity-aligned to better expand the domain knowledge graph and be able to use other relevant entity information in the general knowledge graph to verify the domain knowledge.

[0084] To further implement the above technical solution, as Figure 4 , S2 specifically includes: the large language model extracts the concept words as entities, determines the entity relationships according to the context of the concept words in the current document, extracts triples, and forms a first graph.

[0085] S3 specifically includes: the large language model chains the entities in the first graph to an existing concept graph or a general knowledge graph according to the entity relationships in the first graph;

[0086] Extract the linked entities and entity relationships in the existing concept graph or general knowledge graph to generate a second graph.

[0087] Furthermore, the training steps of the large language model also include:

[0088] Construct an entity disambiguation training set, which includes sample sentences and sample entities in the sample sentences; construct a domain knowledge graph and a general knowledge graph, and generate knowledge graph labels through entity alignment; the knowledge graph labels are provided with target semantic entities and non-target semantic entities corresponding to the sample entities;

[0089] Use the entity disambiguation training set to train the large language model; the trained large language model is used to obtain the target semantic entity corresponding to the sample entity among multiple candidate semantic entities in the knowledge graph to achieve the tasks of triple extraction and entity disambiguation.

[0090] In summary, the outputs of offline training include three parts: the large model used for concept word mining, the traditional model for concept word mining, and the prior knowledge graph data.

[0091] Taking the AI field as an example, the offline construction process is as follows:

[0092] 1. Train the named entity recognition model: Select the large language model baichuan-13B and use the paper abstracts, keywords, and concept words in the openalex dataset as training data. Perform feature engineering, prompt engineering, etc. for development, and perform task fine-tuning on difficult samples in the field so that it can recall the corresponding keywords or concept words for the paper abstract, as well as other vocabulary such as other institutions and people, and give classifications. Then use the fine-tuned baichuan-13B model to predict through external documents such as recent AI news and reports, and obtain the text and the corresponding concept and entity vocabulary, denoted as the "initial training set". Remote supervision can use methods such as sentence pattern template matching and keyword mining to perform cyclic iteration on all news and report data, and finally obtain an extended dataset. Then use this dataset to train the NER model of scrapy as a small model.

[0093] 2. Train the knowledge graph algorithm module for triple extraction and entity disambiguation: First, use the hyponymy relationships of concept words in openalex, relevant concepts of institutions, etc. to construct a domain knowledge graph, and use wikidata data to construct a general knowledge graph. The entities of the two have corresponding relationships and can be directly aligned to generate the graph data of this project.

[0094] For the concept word IDs of the openalex paper abstracts and paper entities, disambiguation training data can be directly constructed; for the "initial training set" in step 1, if a triple in the knowledge graph is matched, the corresponding relationship is established to construct a triple extraction training set. Use a relationship extraction model to perform relationship extraction; the relationship extraction model is a large language model, which inputs a sentence and the two types of entities it contains, and outputs their relationship to form a triple.

[0095] Through development such as feature engineering and prompt engineering, multi-task fine-tuning is also performed on the baichuan-13B model. It should be noted that the prompt words of this model need to point out the current task in a fixed template to accelerate subsequent reasoning.

[0096] 3. Build an offline index for concept relationship extraction and summary: Import the graph data in step 2 into the nebula graph database and the milvus vector database.

[0097] The online reasoning process is as follows:

[0098] 1. Concept word extraction: Input the document, first use the NER model of scrapy to recall the results. If AI domain terms are recalled, then call the fine-tuned baichuan-13B model to confirm again.

[0099] 2. Knowledge Graph Database Building: Use the baichuan-13B model produced by Training Process 2 to obtain the extracted text and disambiguated text from the text using a fixed template and the chain of thought method. The disambiguated text can be recalled using an online search engine such as elasticsearch.

[0100] 3. Concept Word Database Entry: Use the model produced in step 2 to infer triples from the "initial dataset" in step 1 to form a sub- Figure 1 Recall the paths between entities in these triples from nebula to form a sub- Figure 2 The graph similarity algorithm uses the cosine similarity calculated after summing the node vectors in the graph. If the similarity meets the requirements, correct the nodes with the same name but different IDs in the sub- Figure 2 Corrected sub- Figure 1 Then, use the vectors of the entities adjacent to the concept words in the sub- Figure 1 Extract similar entities from milvus and query nebula to obtain connectable paths, and add them to the sub- Figure 2 Generate a sub- Figure 3 . Finally, generate a prompt according to the source document meta-information, the domain definition information of the recalled concept words, the context text of the recalled concept words, and the sub- Figure 3 Paraphrased text. The method for constructing the sub- Figure 3 Paraphrased text can directly concatenate triples and separate them with commas. After arranging the above information in markdown format, ask the baichuan-65B large model one by one and obtain answers. After generating the hyponyms, synonyms, antonyms relationships entities and summaries of the domain words for the answers, write them back to the graph data of this project.

[0101] To further implement the above technical solution, S4 uses the second graph to correct the entity disambiguation result of the first graph. The specific steps of S4 are as follows:

[0102] Calculate the distance similarity between the domain concept words and all entities in the second graph. When there are similar and the corresponding domain concept words are not hit in the first graph, select the corresponding entities of the domain concept words in the second graph to adjust the first graph.

[0103] The adjustment method is as follows:

[0104] Select the nodes with high graph embedding vector similarity in the second graph for recall and add them to the first graph. Generate a third graph according to the known relationships found for the newly added nodes. Specifically, use the graph embedding vectors of all points in the second graph g2, add the top 10 points with a cosine similarity greater than 0.8 to the first graph g1, and find the known paths for all pairs of the newly added points and add them to the first graph g1 to form the third graph g3.

[0105] In another embodiment, to ensure the accuracy of the second graph in S3, S3 further includes: judging the similarity of the generated second graph, and when the similarity meets a preset threshold, executing S4. The calculation method of the similarity is as follows:

[0106] sim = k1 * node_sim(g1, g2) + k2 * edge_sim(g1, g2) + k3 * group_sim(g1, g2)

[0107] Wherein, sim represents the similarity eigenvalue of the first graph g1 and the second graph g2. node_sim is to take out the graph embedding vectors of all the nodes in g1 and g2 and calculate the average value of the cosine similarities respectively; edge_sim is to take out the graph embedding vectors of all the edges in g1 and g2 and calculate the average value of the cosine similarities respectively; group_sim is to take 3-step distances of all the nodes in g1 in the knowledge graph, and the similarity obtained by dividing the number of intersection points of all the nodes and all the nodes in g2 by the number of nodes in g2. k1, k2, and k3 are weight constants. Thus, it is calculated whether the similarity sim of the two subgraphs is greater than the threshold 0.7 to confirm whether the knowledge graph can be recalled. If the similarity is higher than the threshold, then execute S4.

[0108] In another embodiment, S5 is specifically:

[0109] Construct a paraphrased text according to the third graph g3 to represent the relationships between each node, and combine the meta-information of the current document, the known definition information of the domain concept words, and the context text of the domain concept words to form a prompt text; assuming g3 is represented as (s1, p1, o1), (s2, p2, o2),... then the construction method of this paraphrased text is as follows: The {p1's name} of {s1's name} is {o1's name}, the {p2's name} of {s2's name} is {o2's name},...

[0110] When the domain concept word is a new word, confirm the relationship between the new word and the domain words in the library, such as the hyponymy, synonymy, antonymy, etc. relationships, and summarize to generate a definition summary of the new word according to the prompt text.

[0111] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method part.

[0112] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A domain term mining method based on knowledge graph recall, characterized in that, It includes the following steps: S1: Extract domain concept words and their contexts from the current document; S2: Extract triples according to the contexts of the domain concept words in the current document, and generate a first knowledge graph; S3: Link the domain concept words to a prior knowledge graph, and generate a second knowledge graph according to the path relationships of the concept words in the knowledge graph; S4: Select nodes with high similarity of graph embedding vectors in the second knowledge graph for recall, and add them to the first knowledge graph. Find the known relationships in the second knowledge graph according to the newly added nodes to generate a third knowledge graph; S5: Construct a prompt text according to the third knowledge graph to confirm the definition of the domain concept words.

2. The domain term mining method based on knowledge graph recall according to claim 1, wherein The specific content of S1 is: Train a concept word recognition model for extraction. The concept word recognition model includes a named entity recognition model and a large language model; During the offline training process, the large language model is used for vocabulary annotation, generating labeled samples, and recognizing the annotated samples; counting the confidence of each annotated word, and filtering the annotated words through a threshold; used for parameter fine-tuning by developing prompt words according to internal domain documents; Use the large language model to predict and extract high-frequency co-occurring words from external general documents; obtain external annotated samples according to the high-frequency co-occurring words; The named entity recognition model is used for training according to the external annotated samples.

3. The domain term mining method based on knowledge graph recall according to claim 2, wherein, The steps of the concept word recognition model for online model inference include: S11: Conduct a preliminary recall of the current document through the named entity recognition model, and mine document keywords to form a keyword table; S12: Adjust the keyword table through the large language model to obtain the domain concept words.

4. A domain term mining method based on knowledge graph recall according to claim 2, characterized in that The specific content of S2 includes: The large language model extracts the concept words as entities, and confirms entity relationships according to the contexts of the concept words in the current document, extracts triples, and constructs a first knowledge graph.

5. A domain term mining method based on knowledge graph recall according to claim 4, characterized in that, The specific content of S3 includes: The large language model links the entities in the first knowledge graph to an existing concept graph or a general knowledge graph according to the entity relationships in the first knowledge graph; Extract the linked entities and entity relationships in the existing concept graph or general knowledge graph to generate a second knowledge graph.

6. A domain term mining method based on knowledge graph recall according to claim 5, characterized in that, The training steps of the large language model also include: Construct an entity disambiguation training set, which includes sample sentences and sample entities in the sample sentences; construct a domain knowledge graph and a general knowledge graph, and generate knowledge graph labels through entity alignment; the knowledge graph labels are provided with target semantic entities and non-target semantic entities corresponding to the sample entities; Use the entity disambiguation training set to train the large language model; the trained large language model is used to obtain the target semantic entity aligned with the sample entity among multiple candidate semantic entities in the knowledge graph.

7. A method for mining domain terms based on knowledge graph recall according to claim 1, characterized in that, S3 also includes: Judge the similarity of the generated second knowledge graph. When the similarity meets the preset threshold, execute S4.

8. A method for mining domain terms based on knowledge graph recall according to claim 7, characterized in that, The calculation method of the graph similarity is: sim = k1 * node_sim(g1, g2) + k2 * edge_sim(g1, g2) + k3 * group_sim(g1, g2), where sim represents the similarity eigenvalue between the first graph spectrum g1 and the second graph spectrum g2. node_sim is to take out the graph embedding vectors of all the nodes in g1 and g2 and calculate the average value of the cosine similarities respectively; edge_sim is to take out the graph embedding vectors of all the edges in g1 and g2 and calculate the average value of the cosine similarities respectively; group_sim is the similarity obtained by taking 3-step distances of all the nodes in g1 in the knowledge graph, dividing the number of intersection points of all the nodes and all the nodes in g2 by the number of nodes in g2, and k1, k2, and k3 are weight constants.

9. A domain term mining method based on knowledge graph recall according to claim 8, characterized in that, The specific steps of S4 include: Calculating the distance similarity between the domain concept word and all entities in the second graph spectrum; Setting a distance similarity threshold; When the distance similarity threshold is satisfied and the corresponding domain concept word does not hit the first graph spectrum, select the corresponding entity of the domain concept word in the second graph spectrum to recall the first graph spectrum and add it to the first graph spectrum, and generate a third graph spectrum according to the newly added nodes to find the known relationships.

10. A domain term mining method based on knowledge graph recall according to claim 1, characterized in that, The specific step of S5 is: Constructing a paraphrased text according to the third graph spectrum to represent the relationships between various nodes, and combining the meta-information of the current document, the known definition information of the domain concept word, and the context text of the domain concept word to form a prompt text; When the domain concept word is a new word, confirm the relationship between the new word and the domain words in the library according to the prompt text and summarize to generate a definition summary of the new word.

Images